Datasets:
ArXiv:
License:
| # WeatherForecast: Source-resolution data | |
| **English** · [简体中文](https://huggingface.co/datasets/OliverLee520/WeatherForecast/blob/main/Source/README_zh.md) | |
| Pre-downsampling source files for [WeatherForecast](https://huggingface.co/datasets/OliverLee520/WeatherForecast). This is an independently assembled mirror of selected upstream files, not an official release by ECMWF, Copernicus, WeatherBench or the Met2Net authors. “Source resolution” means the input resolution of our downsampling workflow; it does not mean native ERA5 model resolution. | |
| ## Publication status | |
| **Upload complete.** All 48 source files are published. Hub paths and byte sizes match [manifest.json](manifest.json). | |
| ## Relationship to WeatherForecast | |
| | Component | This source directory | Low-resolution counterpart | Dataset task | | |
| |---|---|---|---| | |
| | GlobalForecast | WeatherBench 1.40625°; **128 × 256**; yearly NetCDF, 2010–2018 | [GlobalForecast](https://huggingface.co/datasets/OliverLee520/WeatherForecast/tree/main/GlobalForecast): **32 × 64**, split NPY | GlobalForecast | | |
| | TyphoonTracking | Authors' ERA5-Cropped; **128 × 128**; split NPZ | [TyphoonTracking](https://huggingface.co/datasets/OliverLee520/WeatherForecast/tree/main/TyphoonTracking): **32 × 32**, split NPZ | TyphoonTracking | | |
| No numerical modification is applied to the data files mirrored here. Only repository paths are reorganized. The original compressed download ZIPs and the intermediate 64 × 128 NPY files are not included. | |
| ### GlobalForecast source | |
| - `GlobalForecast/<variable>/<variable>_<year>_1.40625deg.nc`: five variable directories, nine annual files each, 45 files total. | |
| - Variables: `u10` and `v10` (m/s), `t2m` (K), `tcc` (fraction), `r` (%). The relative-humidity source retains **all 13 pressure levels**: 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925 and 1000 hPa. The GlobalForecast derivative uses **only 1000 hPa**. | |
| - Latitude centers −89.296875°…89.296875° (south to north), longitude centers 0°…358.59375°E. Read the actual NetCDF coordinates and timestamps. NetCDF packing is decoded by xarray; do not treat on-disk packed values as physical values. | |
| - Source split frame counts: train 2010–2015: **52,584**; validation 2016: **8,784**; test 2017–2018: **17,520**. Files retain their full annual time axes. | |
| - Derivation: select `r` at 1000 hPa; order channels `u10,v10,t2m,tcc,r`; sample spatial indices `[::2, ::2]` (128 × 256 → 64 × 128); concatenate years per split and remove **only the final frame of each whole split**; store float32; take non-overlapping 2 × 2 means in float64 and store float32 (64 × 128 → 32 × 64). | |
| - The derived time lengths are therefore 52,583 / 8,783 / 17,519. Statistics are recomputed on the derived training set only. This is not a direct 4 × 4 mean and is not the official WeatherBench 5.625° product. | |
| ### TyphoonTracking source | |
| - `TyphoonTracking/train_128_128.npz`: `data.shape=(4,43824,128,128)`. | |
| - `TyphoonTracking/val_128_128.npz` and `test_128_128.npz`: `data.shape=(4,8760,128,128)` each. | |
| - `data` is unnormalized float32 in **C,T,H,W** order: `msl` (Pa), `u10` (m/s), `v10` (m/s), `t2m` (K). `stats` is the upstream object dictionary; use `allow_pickle=True` only when trusting the source and reading this dictionary. | |
| - The reported setup uses hourly frames, 2017–2021 / 2022 / 2023, and a region approximately 110–142°E, 12.25–44°N. These describe the upstream paper/setup: **the NPZ files contain no timestamps or spatial coordinates**, so dates, exact grid endpoints and latitude direction cannot be independently established from these files. | |
| - Derivation: non-overlapping 4 × 4 means in float64, then float32 storage (128 × 128 → 32 × 32), without removing frames. Recompute statistics from the low-resolution training split; do not reuse high-resolution statistics for low-resolution normalization. The mirrored source `stats` are preserved, not newly audited or recomputed. | |
| - Includes non-typhoon periods and **no cyclone-track labels**. These are gridded weather fields, not a track ground-truth dataset. | |
| ## Reading the data | |
| ```python | |
| import xarray as xr | |
| import numpy as np | |
| with xr.open_dataset("Source/GlobalForecast/relative_humidity/relative_humidity_2010_1.40625deg.nc") as ds: | |
| r1000 = ds["r"].sel(level=1000) # decode packed NetCDF values | |
| first_frame = r1000.isel(time=0).values | |
| # Accessing data loads this large compressed NPZ array into memory. | |
| with np.load("Source/TyphoonTracking/test_128_128.npz", allow_pickle=False) as ds: | |
| shape = ds["data"].shape | |
| ``` | |
| For a 12→12 forecasting task, construct windows within each split; do not allow targets to cross split boundaries. Fit normalization only on training data. This source directory does not prescribe the official Met2Net protocol or supply official benchmark scores. | |
| ## Licenses, attribution and redistribution | |
| See [LICENSE.md](https://huggingface.co/datasets/OliverLee520/WeatherForecast/blob/main/LICENSE.md) and [NOTICE.md](https://huggingface.co/datasets/OliverLee520/WeatherForecast/blob/main/NOTICE.md). The repository has **component-specific terms**, not a blanket Apache or MIT relicense: | |
| 1. WeatherBench source data: **CC BY 4.0**, as specified on the [official mediaTUM data record](https://mediatum.ub.tum.de/1524895). The software repository's MIT license is not used as the data license. | |
| 2. Met2Net's source data release: upstream [guaishou1/Met2Net](https://huggingface.co/datasets/guaishou1/Met2Net) declares **Apache-2.0**. Preserve that notice and the included Apache license for this component; underlying ERA5/Copernicus attribution and applicable terms also remain in force. | |
| 3. ERA5: credit Copernicus Climate Change Service and ECMWF, retain applicable data-license notices, and identify modifications. Current CDS entries indicate CC-BY; the earlier Copernicus license also expressly permits redistribution with attribution and disclaimer. Links are provided in LICENSE.md. | |
| Contains modified Copernicus Climate Change Service information (2026 publication; source observation periods 2010–2018 and, as reported by Met2Net, 2017–2023). Neither the European Commission nor ECMWF is responsible for any use that may be made of the Copernicus information or data it contains. Credit upstream authors when redistributing; this mirror claims no ownership of their data, conveys no endorsement, and grants no rights beyond the applicable upstream licenses. Data are provided as-is, without warranty. | |
| ## Sources and citation | |
| - Rasp et al. (2020), *WeatherBench: A benchmark data set for data-driven weather forecasting*. [DOI:10.1029/2020MS002203](https://doi.org/10.1029/2020MS002203); [data](https://mediatum.ub.tum.de/1524895). | |
| - Hersbach et al. (2020), *The ERA5 global reanalysis*. [DOI:10.1002/qj.3803](https://doi.org/10.1002/qj.3803). | |
| - Li et al. (ICCV 2025), *Met2Net*. [Paper](https://arxiv.org/abs/2507.17189); [upstream files](https://huggingface.co/datasets/guaishou1/Met2Net/tree/4995caa72b8e83ea8ff8b77244796af674520f17/era5). | |
| - ERA5 [single levels](https://doi.org/10.24381/cds.adbb2d47) and [pressure levels](https://doi.org/10.24381/cds.bd0915c6). | |
| - For the lower-resolution preparation, cite/link [WeatherForecast](https://huggingface.co/datasets/OliverLee520/WeatherForecast) and state the resolution and processing steps you use. Do not describe either independent release as an official benchmark reproduction. | |